Minimal Oversight — interactive companion

Three browser widgets for "Minimal Oversight: Uncertainty-Aware Governance for Delegated AI Systems" (Azevedo, 2026). Every number is computed live by the paper's equations via mso-core.js, a faithful port of minimal_oversight._formulae pinned to the Python package by a parity test (|Δ| < 1e-6).

Governance cockpit — build & analyze flagship

Build or load a delegated workflow from a connector library (HubSpot, Salesforce, GitHub, Jira, Slack, LLM, human review…), edit the graph, and read feasibility, masking, motifs, and risk off it live.

Feasibility & oversight cockpit

Can a sales pipeline hit its target? Capacity ceiling C_op, the bottleneck, the autonomy buffer, the entropy cliff, and per-node masking — live.

The masking pathology

Why a corrector hides agent weakness. M* = σ_corr/σ_raw, reproducing the paper's M*=1.83.

The Return Operator, run on time

Run the competence ODE forward: σ_raw drifts, σ_corr holds, M* widens. Intervene to reset the autonomy window.

Token simulation (stochastic Petri net)

Run tasks as tokens through the pipeline with real review loops; watch empirical end-to-end success converge to the analytic C_op.

Water-filling oversight

The Euler-Lagrange allocation α*(x) — autonomy where it's cheap, oversight where competence is low — at least cost.

Water-filling vs the baseline paradigm

MSO vs uniform oversight at equal delivery, plus endogenous task allocation (scope selection). The advantage grows with task heterogeneity.

Package: pip install minimal-oversight · github.com/crbazevedo/delegation-lab. To embed a widget, copy mso-core.js + theme.css and the widget HTML, or see web/README.md.